AI Market Developments: The Quarterly Briefing on What's Shifting in Enterprise AI

The three most significant enterprise software platforms — Salesforce (Agentforce), Microsoft (Copilot Agents in M365), and ServiceNow (AI agent integrations) — have all moved agentic AI from preview to production availability. This marks a structural shift: agentic AI is no longer a custom implementation requiring specialist deployment. It is embedded in the platforms enterprises already use.

What this means in practice: procurement teams can deploy AI agents that monitor supplier communications, flag contract renewals, and draft response templates without a dedicated AI implementation project. Operations teams can automate approval workflows, exception flagging, and status reporting within their existing CRM or ERP environment.

The caveat that experienced enterprise AI teams consistently note: embedded platform agents are powerful within the platform's data scope. The more interesting — and more complex — implementations involve agents that cross system boundaries. That is where specialist deployment partners become relevant.

OpenAI's ChatGPT Gov — a secure, government-specific version of ChatGPT that operates within agency-controlled environments on Microsoft Azure Government Cloud — has expanded significantly since its 2025 launch. The platform enables federal agencies to use GPT-4o capabilities while maintaining compliance with security frameworks including IL5, CJIS, ITAR, and FedRAMP High.

Agencies already using the platform include the U.S. Air Force Research Laboratory (administrative tasks, coding), Los Alamos National Laboratory (scientific research), and the Commonwealth of Pennsylvania (project analysis). The significant expansion in 2025-2026 is in defense and intelligence applications where the combination of advanced reasoning capability and classified environment compliance was the primary barrier to adoption.

DeepSeek's arrival in early 2025 demonstrated that high-capability AI models could be built at a fraction of the cost previously assumed. That demonstration accelerated a trend already underway: enterprise AI teams increasingly choosing between three deployment options rather than one.

• Proprietary cloud models (GPT-4o, Claude, Gemini) — highest capability ceiling, usage-based cost, data leaves the organization
• Open source hosted (Meta LLaMA 3.1/3.3, Mistral, Qwen) — strong capability, flexible deployment, cost-efficient at scale
• On-premise open source — full data control, no ongoing API cost, requires infrastructure investment

For manufacturers and B2B operators handling sensitive production data, supplier information, or proprietary process specifications, the on-premise open source option is increasingly viable. The infrastructure cost has dropped significantly as purpose-built local AI hardware has matured.

OpenAI's GPT-5, confirmed by CEO Sam Altman to integrate the O3 reasoning model and be available free to all users at a standard intelligence setting, represents the first time a frontier reasoning model has been broadly accessible without a paid subscription. This removes the cost barrier that previously limited enterprise experimentation with the highest-capability models.

The three capability improvements that matter most for enterprise applications across the major frontier models: significantly improved multi-step reasoning on complex operational problems (not just answering questions but working through ambiguous situations), better instruction-following consistency that makes agentic workflows more reliable, and improved performance on domain-specific knowledge when fine-tuned or given relevant context through RAG.

AI video generation has moved faster from capability demonstration to enterprise deployment than any other generative AI modality. Google's Veo 3 (demonstrated at Google I/O 2025), OpenAI's Sora, Runway Gen-3, and Kling have collectively made text-to-video and video-to-video generation practical for commercial production.

Enterprise applications already in production: marketing content creation (significantly reduced production costs for product demonstration videos), training material development (scenario-based safety and procedure training without location shoots), and customer service visualization (showing rather than explaining complex technical processes).

The manufacturing application is emerging: AI-generated visualizations of process changes, maintenance procedures, and quality control scenarios — reducing the cost and time of updating visual training materials when processes change.

Microsoft's February 2025 announcement of Majorana 1 — the world's first quantum chip using a topological core architecture — generated significant industry attention. The chip leverages topological qubits built on a novel indium arsenide / aluminum materials stack, with eight topological qubits on the initial chip and a design path toward one million qubits.

As Krysta Svore, Microsoft Technical Fellow, described the achievement: "We had to create a new state of matter to reach this point, but once achieved, the architecture itself is elegantly simple and scalable." Chetan Nayak added: "Whatever you're doing in quantum computing, it needs a clear path to a million qubits. Without it, you'll hit a wall before solving the world's most important problems. We've worked out that path."

The realistic 2026 status: Majorana 1 is a significant research achievement and a credible architectural path to practical quantum computing. Commercial applications — including the most compelling ones for manufacturing, like molecular simulation for materials science and complex supply chain optimization — remain years away from practical deployment.

Nvidia's dominance in AI accelerator chips — driven by the H100/H200/Blackwell GPU series and the CUDA software ecosystem — remains the defining fact of AI infrastructure. The company's market capitalization reflects this: as of early 2026, Nvidia is consistently among the three most valuable companies globally, driven by continued enterprise and hyperscaler demand for AI compute.

The CUDA ecosystem is the underappreciated competitive moat. Nvidia's software stack — optimized libraries, developer tools, and the CUDA programming model — means that switching to a competitor chip is not just a hardware change. It requires porting software, retraining teams, and accepting performance uncertainty during transition. This explains why AMD's technically competitive MI300/MI400 accelerators have gained market share without threatening Nvidia's overall position.

US export controls on advanced AI chips to China — restricting Nvidia H100/H800 and AMD MI300 sales — have created two parallel chip markets. Western enterprises have access to the full performance range of advanced AI accelerators. Chinese organizations are limited to domestically developed alternatives, primarily Huawei's Ascend 910B series, and older or restricted versions of Western chips that pre-date control implementation.

The practical implications for global manufacturers: organizations with operations in China face a bifurcated AI infrastructure decision — cloud-based AI services using available domestic providers (Alibaba Cloud, Baidu AI Cloud, Huawei Cloud) or on-premise infrastructure using domestically available hardware. Neither offers the performance parity with Western infrastructure that was available before controls.

DeepSeek's January 2025 launch was genuinely disruptive — a Chinese open-source model built in two months at under $6 million that matched GPT-4 class performance on reasoning benchmarks, challenging the assumption that frontier AI required billion-dollar training budgets. Mark Zuckerberg assembled four engineering war rooms to analyze how it was done. Nvidia lost $630 billion in market cap in a single day on the news.

The disruption was real. Its durability is more complicated. Multiple governments banned DeepSeek from government devices (US, South Korea, Australia, Italy, Taiwan) following documented evidence of data transmission to Chinese servers despite privacy claims — an independent investigation by IT researcher David Bombal found 51 data packets transmitted to Beijing infrastructure. Perplexity moved quickly to offer a DeepSeek R1-based service hosted entirely in US and European data centers, addressing the data sovereignty concern.

The more lasting impact of DeepSeek is not the model itself but what it demonstrated: the efficiency frontier for AI training had moved dramatically. US labs responded by intensifying their own efficiency research. The $6M training cost of DeepSeek-R1 shifted the conversation about what AI development could cost — and by implication, who could compete.

China filed over 38,000 patents for generative AI technologies between 2014 and 2023 — approximately 70% of the world's total. The United States filed 6,276 in the same period. In 2024 alone, China filed 300,510 AI-related patents, compared to 67,773 by the US.

The quality adjustment changes the picture significantly. US AI patents are cited nearly seven times more often than Chinese counterparts — an average of 13.18 citations per US patent versus 1.90 for Chinese patents. This suggests that while China leads in volume, the most foundational and broadly applicable AI research remains predominantly American.

The practical reading: China has built a massive AI application and deployment capability on top of research foundations that remain largely Western. DeepSeek's breakthrough was architectural efficiency, not fundamental research — it found a more efficient way to train on existing architectures. That is a significant capability, but different from the foundational research that creates new architectural possibilities.

The European Union's InvestAI initiative — announced at the AI Action Summit in Paris with €200 billion in combined public-private investment — represents Europe's most ambitious attempt to establish AI competitive parity with the US and China. The initiative includes AI gigafactories (large-scale shared computing hubs modeled on the CERN research model), €150 billion in private investment, and €50 billion in EU public funding.

Europe's approach is architecturally different from both US and Chinese strategies: open-source development, shared infrastructure, and safety-first regulation through the EU AI Act. The theory is that by avoiding the proprietary model development race, Europe can build AI capability without trying to out-invest hyperscalers on a per-company basis.

The persistent challenge is the brain drain problem — European AI talent continues to flow toward US tech companies and research institutions at salaries that EU organizations struggle to match. InvestAI's human capital components address this explicitly, but the structural salary differential is not resolved by policy announcements.